Automatic Mapping Clinical Notes to Medical - RMIT University

Named Entity Recognition for Astronomy Literature
Tara Murphy and Tara McIntosh and James R. Curran
School of Information Technologies
University of Sydney
NSW 2006, Australia
{tm,tara,james}@it.usyd.edu.au
Abstract
We present a system for named entity
recognition (ner) in astronomy journal
articles. We have developed this
system on a ne corpus comprising approximately
200,000 words of text from
astronomy articles. These have been
manually annotated with ∼40 entity
types of interest to astronomers.
We report on the challenges involved
in extracting the corpus, defining entity
classes and annotating scientific
text. We investigate which features of
an existing state-of-the-art Maximum
Entropy approach perform well on astronomy
text. Our system achieves an
F-score of 87.8%.
1 Introduction
Named entity recognition (ner) involves assigning
broad semantic categories to entity references
in text. While many of these categories
do in fact refer to named entities, e.g.
person and location, others are not proper
nouns, e.g. date and money. However, they
are all syntactically and/or semantically distinct
and play a key role in Information Extraction
(ie). ner is also a key component of
Question Answering (qa) systems (Hirschman
and Gaizauskas, 2001). State-of-the-art qa
systems often have custom-built ner components
with finer-grained categories than existing
corpora (Harabagiu et al., 2000). For ie
and qa systems, generalising entity references
to broad semantic categories allows shallow extraction
techniques to identify entities of interest
and the relationships between them.
Another recent trend is to move beyond the
traditional domain of newspaper text to other
corpora. In particular, there is increasing interest
in extracting information from scientific
documents, such as journal articles, especially
in biomedicine (Hirschman et al., 2002).
A key step in this process is understanding
the entities of interest to scientists and building
models to identify them in text. Unfortunately,
existing models of language perform
very badly on scientific text even for the categories
which map directly between science and
newswire, e.g. person. Scientific entities often
have more distinctive orthographic structure
which is not exploited by existing models.
In this work we identify entities within astronomical
journal articles. The astronomy
domain has several advantages: firstly, it is
representative of the physical sciences; secondly,
the majority of papers are freely available
in a format that is relatively easy to manipulate
(L ATEX); thirdly, there are many interesting
entity types to consider annotating;
finally, there are many databases of astronomical
objects that we will eventually exploit as
gazetteer information.
After reviewing comparable named entity
corpora, we discuss aspects of astronomy that
make it challenging for nlp. We then describe
the corpus collection and extraction process,
define the named entity categories and present
some examples of interesting cases of ambiguity
that come up in astronomical text.
Finally, we describe experiments with retraining
an existing Maximum Entropy tagger
for astronomical named entities. Interestingly,
some feature types that work well for
newswire significantly degrade accuracy here.
We also use the tagger to detect errors and
inconsistencies in the annotated corpus. We
plan to develop a much larger freely available
astronomy ne corpus based on our experience
described here.
Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 57–64.
57